Anubrat Sahoo — AI/ML & Computer Vision Engineer
AI/ML & Computer Vision Engineer
About
I build robust artificial intelligence pipelines and scalable web applications. My focus is on Computer Vision, leveraging models like PyTorch, RF-DETR, and DINOv2 for complex OCR and object detection tasks, alongside building conformal calibration systems for high-reliability outputs.
Featured Case Study
The Illusion of Solved OCR
Everyone thinks OCR is a solved problem. With the release of GPT-5, Claude Sonnet 5, and Gemini 3.1 Pro, the consensus is that Vision-Language Models (VLMs) have conquered text extraction. They haven't. They crush synthetic benchmarks, but they fail silently and dangerously in the real world. Generalist VLMs are trained to reason, solve, and infer. But in production OCR (like grading student exams or processing medical records), reasoning is a liability. You need verbatim extraction.
The Problem: Intelligence vs. Extraction
SOTA models fail at real-world extraction because they are trained to reason, not to faithfully transcribe.
Failure A: The 'Hallucination of Intent'
Generalist models are 'too smart for their own good.' When a student implies a step but doesn't explicitly write it, models like Gemini hallucinate the missing text into the transcript.
"The intention of the guy was to divide 80 by 2 but he never wrote 80/2. Gemini wrote it in its transcript. I feel like this is a very risky thing."
Failure B: Handling Real-World Noise
When confronted with scribbles, crossing-out, or ambiguous handwriting, massive VLMs get confused and pollute the transcription data.
"2nd step 4n (the kid meant step 1) but Gemini transcript shows this as -1... Gemini 3.1 Pro was hella confused about what it exactly transcribes so this proves that all these synthetic benchmarks do not reflect real-world performance."
Building Narrow VLMs to Beat SOTA
My approach: Build a pipeline of highly specialized, narrow models (like a chemistry-exam VLM trained on just 1,500 images paired with an RF-DETR object detector) that completely disappears into the product. It beats SOTA models like Claude Sonnet 5 at a fraction of the cost.
Phase 1: Identifying the Noise (Object Detection)
The OCR model shouldn't even try to read crossed-out words. I picked up the RF-DETR architecture and trained a custom model from scratch specifically to recognize scribbles, small cuts, and large strikethroughs performed by students.
Phase 3: Outperforming SOTA
By restricting the domain and sanitizing the inputs, I trained a narrow chemistry-exam VLM on just 1,500 images. This specialized engine punches above its weight class, beating Claude Sonnet 5 on specific extraction tasks while being small enough to run cheaply and fast in production.
Selected AI Work
AI & Machine Learning
A multimodal AI workspace routing between Qwen-VL, InternVL, DeepSeek, and GLM by task type, with visual output generation (interactive math graphs, 3D molecular structures, DFA/NFA diagrams) and a multi-step research mode over Tavily.
Engineered a robust computer vision pipeline for automated grading and benchmarking. Leveraged OCR, RF-DETR, and DINOv2 for high-accuracy document parsing, and implemented conformal calibration for reliable confidence scores.
Building reliable systems for LLM agents through secure sandboxed execution environments, robust tool-calling loops, and scalable production backend infrastructure.
Built a platform that evaluates job readiness by pulling resume data and GitHub contribution activity, then runs scoring logic to estimate market-aligned salary ranges. Deployed on AWS Fargate with PostgreSQL and S3.
Supporting Experience
Web3 & Blockchain
Full-stack Web3 application with Solidity contracts deployed on mainnet. Smart contracts are the system of record; frontend logic is fully constrained by on-chain state. Contract logic tested with Foundry. Received ecosystem grant support.
Modular protocol core using the Diamond Standard — facets for access control and execution logic, strict storage layout conventions to prevent slot collisions across upgrades.
Smart contracts reasoning about state and interactions across L3 and L2 layers. Focused on cross-layer assumptions, state consistency, and correct execution semantics.
Open Source
Recent Pull Requests
Writing & Explainers
Currently Exploring
- Vision-Language Models (VLMs) and their integration into agentic workflows
- Evaluating OCR extraction reliability for structured document parsing
- Calibration techniques for detection models like RF-DETR and DINOv2
Skills
Languages
Python, TypeScript, JavaScript, Solidity, SQL
AI & Machine Learning
PyTorch, Computer Vision, OCR, RF-DETR, DINOv2, VLM
Frontend
React, Next.js, Tailwind CSS
Backend
Node.js, Express, REST APIs, PostgreSQL
Cloud & Infra
Docker, AWS (S3, RDS, Fargate), CI/CD
Web3
Solidity, EVM, Foundry, Smart Contracts
Quick Answers
Yes. Anubrat Sahoo is actively open to work and looking for AI/ML engineering, Computer Vision, or full-stack engineering roles. He can be reached at anubrat23@gmail.com.
Anubrat Sahoo specializes in Computer Vision and works with PyTorch, OCR, RF-DETR, DINOv2, conformal calibration, and Vision-Language Models (VLMs). He builds robust benchmark pipelines like IntelGrader.
As a strong full-stack backup, Anubrat Sahoo works with React, Next.js, Node.js, TypeScript, PostgreSQL, REST APIs, and AWS (S3, RDS, Fargate). He has shipped production web applications covering backend APIs and cloud deployments.
Anubrat Sahoo has built and deployed Solidity smart contracts to mainnet, including a grant-backed full-stack dApp. He works with Foundry and upgradeable contracts. Web3 is a supporting credibility in his engineering repertoire.
Anubrat Sahoo is open to AI/ML engineering, Computer Vision, and full-stack positions. He is a third-year CS student at SOA University, Bhubaneswar, with production experience and extensive AI projects.
Reach out directly at anubrat23@gmail.com or connect on LinkedIn at linkedin.com/in/anubrat-sahoo. He is actively looking for AI/ML or full-stack roles and responds quickly to serious enquiries.
Immediately available. Open to full-time, part-time, contract, or internship roles in AI/ML, Computer Vision, or full-stack engineering.